A Unified Model for L1 and L2
Brian MacWhinney
HKIEd, Carnegie Mellon
Unified Model
Thanks to ...
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Elizabeth Bates
Csaba Pléh
Julia Evans
Ping Li
Yoshinori Sasaki
Reinhold Kliegl
Vera Kempe
Elena Pizzuto
Roman Taraban
Yuki Yoshimura
Chris Jones
Yvan Rose
Nora Presson
Yanhui Zhang
• NIMH (25 years)
Michèle Kail
Klaus Köpcke
Natasha Tokowicz
Igor Farkas
Richard Wong
Jeff Sokolov
Janet McDonald
Stan Smith
Patricia Brooks
Melita Kovacevic
Jared Leinbach
Kees De Bot
Yanping Dong
Sue-mei Wu
Kerry Kilborn
Maryellen MacDonald
Ovid Tzeng
Arturo Hernandez
Antonella Devescovi
Beverly Wulfeck
Hasan Taman
Dan Slobin
Zhou Jing
Joe Stemberger
Christophe Parisse
Phil Pavlik
Anat Prior
NSF (10 years) MacArthur (3 years)
Unified Model
Economic Assumptions
• Competence in English is crucial for success in the
global economy.
• But most of the population of the world does not
speak English as L1. So English is L2. Other L2s
have parallel roles.
• It is not enough to restrict L2 competence to the elite,
since work is becoming increasingly based on
language skills.
• Different social and economic configurations will
require differing levels of L2 competence.
Unified Model
Position 1: Early Immersion
• There is a Critical Period for language
learning.
• There is a learning/acquisition dichotomy. Late
bilinguals can never achieve full L2
competence.
• Therefore, we must start immersion L2
programs at the pre-primary level.
• And spend billions of dollars in exposure, but
not really teaching. Unified Model
Position 2: Focus on community
• There is a Critical Period and a
learning/acquisition dichotomy.
• However, immersion will not work and can
conflict with other goals in early childhood
education.
• Pre-college education should be in the native
language.
• Full bilingualism is only possible if the
community becomes bilingual.
Unified Model
Position 3: Focus on quality
• There is no critical period for second language
learning, although there are important age effects.
• Critical period effects are due to entrenchment and
competition.
• What is important is not the timing of learning, but the
quality of exposure.
• We may still need billions of dollars, but in teaching,
not just exposure.
• Languages can be learnt and taught. There is no real
learning/acquisition dichotomy.
Unified Model
The Positions
• Position 1 -- UG: Chomsky, Lenneberg,
Krashen, Long, Hurford, Pinker, Newport,
Meisel
• Position 2 -- Sociolinguistics: Fishman, Swain,
Ervin-Tripp, Gumperz
• Position 3 -- Emergentism: Bates, Ellis,
Bialystok, Snow, MacWhinney, Ringbom
Unified Model
7 Pillars of UG
1. Critical Period -- today’s focus
2. Grammar Gene
3. Speech is Special
4. Modularity
5. Poverty of the Stimulus
6. Sudden Evolution of Language
7. Centrality of Recursion
Unified Model
7 pillars of emergentism
1.
2.
3.
4.
5.
6.
7.
L1-L2 competition and entrenchment
Gradual evolution
Modules are made not born
Polygenic emergent genome
Speech relies on mammalian abilities
Learning on input
Emergence of recursion
Unified Model
Which will stand?
Unified Model
Entrenchment vs. Critical Periods
•
•
•
•
•
•
•
Critical Periods are linked to infancy.
Observed drop is not precipitous.
Lateralization is not linked to CP.
Language is not a unitary ability.
Golf, ballet are also age-related.
No mechanism has been discovered.
UG-related syntactic patterns are not strongly
fossilized - Birdsong
Unified Model
Critical Periods
• Bee dance, cricket song
• Does the ability need a trigger?
• When does it start and end?
Unified Model
12
L1 CP≠ L2 CP
L’enfant Sauvage by
François Truffaut
Truffaut as Dr. Jean Itard
Unified Model
13
How many CPs?
•
•
•
•
•
•
•
•
6 mos -- deaf children
2 -- Early bilingual impacts
5 -- Output phonology Flege
8 -- Korean adoptees, literacy, orthography
13 -- Hemispherectomies, synaptic pruning
15 -- Shift in learning, growth of strategies
20 -- Beginning of decline
40 -- Social difficulties
Unified Model
14
Where is the critical drop?
• Newport & Johnson
Unified Model
Hakuta actual
15
A real CP - Hubel & Weisel
P E R C E N T L E F T V IS U A L C O R T E X C E L L S
R E S P O N D IN G T O C L O S E D R IG H T E Y E
(N o rm a l > 5 0 % )
A g e a t e y e c lo s u r e ( d a y s )
D u ra tio n
10
23
30
60
120
1 8 0 A d u lt
(d a y s )
0
84
3
32
6
0
9
0
21
26
27
14
30
19
31
12
65
10
90
120
480
59
15
53
70
Unified Model
16
What we know
• Critical periods are basic to embryology.
• Critical periods for binocular vision in cats; periods for
exposure to song in birds; precocial bird attachment;
• Animals have many instincts; but is language an
instinct?
• Kuhl and Werker: brain locks in on early sounds
• Bosch, Juszyck: Auditory system builds early contrasts
• Rosenzweig rats in rich environments get bigger brains.
Unified Model
17
A bridge too far
•
•
•
•
•
•
No evidence for early brain effects
Mozart for babies
Linda Acredolo and Baby Signs
Mobiles, language while you sleep
Suzuki method
There is nothing wrong with early L2 learning,
but no evidence that it is indispensable
• Early bilingualism ≠ Early L2 learning
Unified Model
18
CP for holding pens?
Unified Model
Chopsticks?
Unified Model
Multiple language abilities
• Bulgarian grad student who wrote at the top of
the class, but had a noticeable accent.
• Hungarian diplomat with perfect English, but
nothing to say.
• Japanese grad student with perfect interaction
and comprehension, but impossible definite
articles and slow test-taking.
• Fossilization for specific German nouns vs.
fossilization for some past tenses.
Unified Model
How can we decide?
• Neurological evidence for a Critical Period
• Immigrant studies


Proof of success in native acquisition for age of arrival well past the
Critical Period.
Proof of failure after some early age of arrival.
• L2 Classroom studies



Big correlational analyses (questionable method)
Randomized clinical trials (if we could get funding)
Microgenetic method studies (my current preference)
 experiments -- can we teach r/l?
 online methods
 TalkBank video methods
Unified Model
Mechanisms of UG
• Genes
• Modules
• Principles, Parameters, Rules
Unified Model
Mechanisms of Emergence
•
•
•
•
•
•
•
Entrainment, physical and social
Adaptation, selection
Competition, strength, reinforcement
Maps, topology, short connections
Self-organized criticality
Resonance
Homeostasis, homeorhesis, feedback
Unified Model
Why the shift to emergentism?
•
Without advanced methods, emergentist cognitive
science was not possible
•
We didn’t have CHILDES, TalkBank
•
Audio, video analysis was primitive
•
We couldn’t simulate - PDP, SOM, ART
•
We couldn’t image the brain - ERP, fMRI
•
We couldn’t study learning in vivo - PSLC.
•
With these advances, emergentism is becoming the
default stance.
Unified Model
Unified Competition Model
maps
transfer
chunking
buffers
resonance
competition
codes
mental models
Unified Model
L1 and L2
• The learning goals are the same.
• The available mental processes are the same.
• However, the specific challenges are different.
Unified Model
27
L1 Learning Challenges
•
•
•
•
Segmenting out words
Organizing phonological gestures
Bootstrapping syntax
Conversational sequencing
Unified Model
28
L2 Challenges
• Maximizing positive transfer
• Avoiding negative transfer
• Overcoming age effects


Using resonance to overcome entrenchment
Proceduralizing declarative structures Ullman/Paradis
Unified Model
29
Component Theories
1. Competition
2.
3.
4.
5.
6.
7.
8.
interactive activation, Bayes
Maps
SOM, entrenchment
Transfer
A relation between maps
Chunking
chunking theory, fluency
Buffers
processing load, CAPS
Resonance
memory theory, Pimsleur, coding
Mental model perspective, embodiment
Codes
sociolinguistics, identification
Unified Model
1. Cue Competition
•







Whodunit?
The tiger pushes the bear.
The bear the tiger pushes.
Pushes the tiger the bear.
The dogs the eraser push.
The dogs the eraser pushes.
The cat push the dogs.
Il gatto spingono i cani.
Unified Model
Cues vary across languages
• English: The pig loves the farmer
SV > VO > Agreement
• German: Das Schwein liebt den Bauer.

Den Bauer liebt das Schwein
Case > Agreement > Animacy>Word Order
• Spanish: El cerdo quiere al campesino.

Al campesino le quiere el cerdo.
"Case" > Agreement > Clitic > Animacy > Word Order
Unified Model
Cues
Device
Example
Word Order
the dog chases the cat
Function words
der - die - das
Affixes
was tak-en
Clitics
nous, le, ba
Constructions
the more -- the merrier
Unified Model
Central Claim
•Cue validity predicts cue strength
•(Bayesian statistics)
[p(function)|form] - comprehension

[p(form)|function] - production

Cue validity measured in corpora

Cue strength measured in experiments

Unified Model
Cues Compete
The bear the tigers chases.
“Tigers”-as-Agent
preverbal position
competes
“Bear”-as-Agent
SV agreement
Unified Model
Initial Position
L1/L2 Competition
I often go ... / Je vais souvent ...
V + Adv
Adv + V
competes
speaking English:
speaking French:
ADV 1st
Unified Model
ADV 2nd
Heavy Adv
Strength measured in experiments
Unified Model
English Children
Unified Model
Hungarian Children
Unified Model
Italian Children
Unified Model
English L1, Dutch L2
Dissertations by Janet McDonald and Kerry Kilborn
Unified Model
Dutch L1, English L2
Unified Model
Findings - 22 studies
• Validity predicts Strength.
• Children and L2 learners pick up frequent cues
first, then they settle on reliable cues.
• For timed tasks, strong fast cues dominate.
• L2 learners attempt transfer, but then learn
cues, as in L1. They gradually reach L1 levels
of cue strength.
Unified Model
2. Maps
•
•
Maps are central to the processing theory. They
control transfer, entrenchment, and embodied
encoding.
Maps are emergent:
- Neural systems: Jacobs & Jordan 1992
- Children: Karmiloff-Smith 1997
- Robots: Nolfi 1996, Tani 2002
Unified Model
Self-organizing lexical maps
Li, Farkas, MacWhinney
- Neural network
- computer simulation
- L1 lexical learning
- CHILDES input
- no initial organization
- short connections
Unified Model
Gradual Emergence
50, 150, 250, 500 words
Unified Model
Refining competition
Unified Model
Bilingual self-organization
Word
Form
Phonological Map
Phonological
ENGLISH PHONOLOGY
Self-organization
CHINESE PHONOLOGY
Chinese Phonology
ASSOCIATIVE
CONNECTIONS (Hebbian learning)
Word
Meaning
Co-occurrence-based
representation
(derived from separate component exposed to bilingual
corpus)
Self-organization
CHCHINESE SEMANTICS
Chinese
Semantics
ENGLISH SEMANTICS
Semantic Map
Unified Model
Maps implement entrenchment
• Strong items dominate over weak.
• Late L2 items are parasitic on pre-existing L1
forms and maps
Unified Model
Module Entrenchment
Simultaneous Bilingualism
LX
LY
balanced
Successive Bilingualism
L1
L2
dominates
Unified Model
3. Transfer
• Mapability
 Item-based (want X) patterns will not transfer
 Grammatical semantics can be a difficult map
 Phonology, semantics, pragmatics all map and transfer
• Markedness
 Unmarked pattern-based will: Adv + V
 Marked pattern-based is weak: Adv + V + S
 Semantic/phonological prototypes transfer
• Filtering
 Japanese r/l second formant transitions.
 English learners of tones.
Unified Model
Examples
•
•
•
•
taco -> t’aco
wenn (if) -> when
tell me a story -> say me a story
install a new version -> install new version
Unified Model
The Culprits
•
•
•
•
•
•
Entrenchment
Transfer (crosstalk)
Learning your own errors
Strategy blockage
Social culprits
Aging
Unified Model
Social Culprits
• Overcommitment
too much email, too many committees

• Declining L2 contact environment
• Avoidance of L2 input
• Allegiance to L1
Unified Model
Aging
•
•
•
•
•
•
Loss of Auditory Acuity - age effects
Loss of Motor Control - Parkinsonism
Cell death -- both cortical and white matter
Declining transmission speed
Declining hippocampal storage
Trauma
Unified Model
Fighting back
1.
2.
3.
4.
Undoing transfer
Unblocking social barriers
Unblocking strategy barriers
Increasing differentiation and resonance
Unified Model
Overcoming Parasitism
C
L1
turtle
C
L2
L1
tortuga
turtle
Unified Model
L2
tortuga
ERP evidence of transfer - P600
• The cat likes to eat. vs The cat likes to eating.
P
z
P600
5
V
Plausible (eat)
Implausible (eating)
200
400
600
800
Osterhout & Nicol (1999)
Unified Model
L1 supports L2
Su abuela cocina/*cocinando muy bien.
Her aunt cooks/*cooking very well.
Unified Model
L1 (English) blocks attention
El/los libros son muy interestantes.
The/the books are very interesting.
Unified Model
L2 cares, L1 oblivious
Ellos fueron a una/*un fiesta.
They went to a/a party.
Unified Model
Behavioral Data
Unified Model
4. Chunking
•Task: Repeat 坐公共汽車去
•Learn: gōnggòngqìchē “bus” 公共汽車
Syllables plus tone encodings fill working memory
Chunk: gōnggòng is linked to “public”
Chunk: qìchē is linked to “motor car”
Supportive links to characters
Compound is a weak chunk, weak tone sequence
Embed weak chunk in “sit ___ go” frame 坐 (公共汽車) 去

Unified Model
Translation Disfluency
•
Do you want to take a bus to Nanjing next week?
•
Nǐ xiǎng xià ge xīngqī zuò gōnggòngqìchē qù Nánjīng
ma?
•
Chinese requires temporal before verb.
•
About to say: Nǐ xiǎng zuò
•
•
•
•
Pause ….
Insert “xià ge xīngqī”
Continue
Result: Non-fluency
Unified Model
Chunks mesh into slots
•
•
•
sit + (vehicle slot) + go
(adverb slot) + V
(topic slot) + comment
•
Fluent plan emerges from coordination of
individual item-based patterns
Unified Model
PSLC studies of Fluency
•
•
Online Dictation -- French, Chinese
Yuki Yoshimura’s CMU dissertation on
Fluency in Japanese L2 - sentence repetition
after reading and listening.
Unified Model
Repetition and WM
16
C o m p a ris o n b e tw e e n re a d - a lo u d a n d p ro d u c tio n tim e
C o m p a ris o n b e tw e e n re a d - a lo u d a n d p ro d u c tio n tim e
C o m p le xit y = sim p le
C o m p le x it y = c o m p le x
rea d -alou d
p rod u c tion
16
14
L e n g th o f u tte ra n ce (s e c)
Length of utterance (sec)
14
12
12
10
10
8
6
4
2
0
read -alou d
p rod u ction
4
6
8
10
8
6
4
2
0
S en te n c e L e n g th
4
6
8
S en ten ce L en g th
Unified Model
10
Omissions
E rro r An alysis b y typ e
co m p lexity = co m p lex
1
N um ber of e rrors in production
o m is s io n
0 .8
re t ra c e
0 .6
g ra m m a t ic a l
e rro r
0 .4
s ub s t it ut io n
0 .2
a dd it io n
0
4
6
8
S en ten ce L en g th
Unified Model
10
Adding Novel Words
E ffe c t o f n o v e l w o rd s
L is te n in g g ro u p x R e a d in g g ro u p
1 60 00
Lis te ning
L en g th o f u ttera n ce (m s)
Re a ding
1 40 00
1 20 00
1 00 00
8 00 0
6 00 0
ze ro
1
2
N u m b e r o f n o v e l w o rd s in e a c h s e n te n c e
Unified Model
Friederici
• German Natives show

for semantic violations: N400

for syntactic violations: ELAN & P600
• L2 Russian natives - 5 years in Germany

for semantic violations: N400

for syntactic violations: no ELAN, but P600
• Brocanto and mini-Nihongo Learners: ELAN and P600
• fMRI Conclusion: L1 and L2 use same areas, but L2
relies more on Broca’s
Unified Model
5. Buffers
• Competition occurs in buffers
• Incrementalism, role-slot filling
• This is developed in




MacWhinney (1987)
Kempen & Hoenkamp (1987)
Levelt (1990)
O’Grady (2006)
Unified Model
71
6. Resonance
• Graduated interval recall
• Multimodal consolidation
• Self-organized criticality
Unified Model
Graduated interval recall
•Pimsleur 67
Unified Model
Neural Basis
Wittenburg et al. 2002
Unified Model
Consolidation Circuits
Dynamic
Meaning
Sound
Consolidation
Hippo
campus
Basal
Ganglia
Scaffold
Unified Model
Chinese Resonance
Unified Model
Consolidation and Time
• Bones, muscles, cell walls, mitochondria, and immune
system becomes stronger after periods of use and
breakage.
• These systems respond to pressures across time frames.
(slow muscles, fast muscles)
• Neurons work the same way.
Unified Model
Math Models: Pavlik 2006
t=time from practice
d=decay rate
n=number of presentations
m=memory activation
a=base decay rate
c=scales effect of activation
on decay
u=maximal study benefit
v=rise to asymptote speed
Unified Model
Four Pools
• Pool 1 – item is strong, then wait
• Pool 2 – item is weak enough to make practice efficient but
strong enough to make drilling more efficient
• Pool 3 – item is weak and retrieval will fail, so study
practice is more efficient
• Pool 4 – unpracticed items
• Algorithm selects items in this order: 2, 3, 4, 1
• Learned items are removed from pools
Unified Model
Optimization really helps
Unified Model
7. Mental models
•
•
•
•
We build up mental models through
perspective-taking.
Comprehensible input -- L2 speaker can
construct a coherent mental model.
L2 conversation-based teaching has to make
sure the mental model is on track.
Frames, scaffolds, can support this.
Unified Model
8. Codes
• Code-switching
• L2 is a code choice
• Codes involve perspective taking in mental
models
• Role of video in learning, identification
Unified Model
82
The Unified Model
•
•
•
•
•
•
•
Competition is central.
Both L1 and L2 are emergent.
Item-based constructions compete in L1 and L2
learning.
Transfer arises from entrenchment in maps.
Fluency develops through chunk meshing.
Resonance and spacing produce robust learning.
Conversation supports perspective switching and
model construction.
Unified Model
Conclusions
•
•
•
•
The Unified Model integrates our
understanding of first and second language
acquisition.
Language learning relies on emergentist
processes.
Language can be taught and learned.
Age-related effects arise from entrenchment
and social commitment, not UG.
Unified Model
Links
• http://psyling.psy.cmu.edu/papers
• http://psyling.psy.cmu.edu/talks
Unified Model
85
Aphasics - Word Order
Unified Model
Aphasics - Agreement
Unified Model
Descargar

CHILDES - Brian MacWhinney's Home Page